Pig sounds can reflect physiological information and health of pig.It is helpful for early detection of the sows rut, preventing the disease and avoiding stress.Identification of pigs different sounds can improve the automation of animal husbandry, information level and animal welfare.Sound recognition system has the characteristics of non-contact and continuity.And it can reduce the labor intensity of workers.This technology has broad application prospects.The purpose of this paper is to identify three kinds of pig barking, namely estrus sounds, cough, howling.Automatic feature extraction and classification of three kinds of sounds are the focus of the study.Because of the complexity of the environment, the paper remove noises based on the improved wavelet de-noising theory firstly, and then carry on the endpoint detection in the preprocessing module.Then, the formants, linear prediction cepstrum coefficients (LPCC), Mel frequency cepstrum coefficients (MFCC) etc.feature parameters are extracted and simulation in the feature extraction module.Through analysis and comparison, the MFCC is as parameters of the main feature parameters of the sound recognition.In this paper, we propose an improved feature parameter MFCC_B, which takes into account the characteristics of the pigs calls, and describes the characteristics of the call signal more accurately and more comprehensively.And finally select MFCC_B as the characteristic parameters of sound recognition.The DTW, HMM and GMM recognition models are trained by MFCC_B features, and the GMM model is better than the other two algorithms by comparing and analyzing the results.In this paper, we propose an improved feature parameter GMM MFCC_B recognition method which has low dimension and low computational complexity, and the recognition rate is above 95%.